Build a Private AI Knowledge Base with Webman AI, Redis‑Stack, and Ollama

Overview

Webman AI 5.4.0 introduces support for arbitrary embedding models, allowing seamless integration with local open‑source models such as DeepSeek to create a private AI knowledge base that operates without external network connections.

Training Module Deployment Steps

Step 1: Install Redis‑Stack Server

Note: A standard redis‑server does not support the required features; you must use redis‑stack.

docker pull docker.ms.run/redis/redis-stack
mkdir /home/data/redis -p
docker run --name redis-stack -v /home/data/redis:/data -p 6380:6379 -d redis/redis-stack

Step 2: Install illuminate/redis Component for Webman

Note: For Webman v2 run the command in the Webman root directory.

composer require -W webman/redis illuminate/events

Note: For Webman v1 run the command in the Webman root directory.

composer require -W illuminate/redis illuminate/events

Step 3: Configure Redis

Create plugin/ai/config/redis.php with the following content:

<?php
return [
    'default' => [
        'host' => '127.0.0.1',
        'password' => null,
        'port' => 6380,
        'database' => 0,
    ],
];

Install Ollama and Add DeepSeek and Other Models

Install Ollama and pull the required models:

ollama pull nomic-embed-text
ollama pull deepseek-r1:14b

Tip: nomic-embed-text is a vector model required for training. Besides deepseek-r1:14b , you can also choose models such as deepseek-coder-v2:latest or qwen2.5:14b .

Step 4: Restart Webman

Linux: php start.php restart -d Windows:

ctl c   # stop webman
php windows.php start

Training Process

Add Training Dataset

Upload your training files (supported formats: .txt and .md, each file ≤ 9 KB; compressed archives are also accepted with the same size limit per file).

Tip: If you see "Class 'Redis' not found", install the PHP Redis extension.

Import Data

Tip: Training data has no fixed format; prepend each entry with a concise title or question to improve results.

Start Training

When the vector field of a record contains a value, the entry has been successfully trained and its vector is stored in Redis‑Stack for fast retrieval.

Set Role Prompt

In the role configuration, select the training set and save (frontend users need to refresh the page for the role to take effect). The AI will then answer queries using the trained data.

Test Results

With the role prompt such as "Webman is a high‑performance PHP framework; you are a Webman assistant. Answer based on the Webman documentation, and if the question is unrelated, reply that no corresponding answer was found," the AI provides accurate, documentation‑based responses.